Hi Ricky,
If all 4 observations per subject have the same pairwise covariances, then they are exchangeable, and you can use randomise with the option "-e design.grp" to supply the exchangeability blocks (EB). In PALM the same is done with "-eb design.grp".
If the above assumption about the covariance within-subject isn't met, but the errors are symmetric around zero (typically they are for FMRI data), then you can use whole-block sign-flipping: in randomise use the options "-e design.grp --permuteBlocks -1", or in PALM the options "-eb design.grp -whole -ise".
More details about this are in the randomise
paper.
Hope this helps.
All the best,
Anderson